Research On Staging Method Of Diabetic Retinopathy Based On Deep Learning

Diabetic Retinopathy(DR)is one of the most serious complications of diabetes.At present,DR detection mainly relies on detailed analysis of fundus images by ophthalmologists,so as to formulate different treatment plans according to the detection results of different patients.However,manual diagnosis is time-consuming and inefficient.The rapidly growing number of patients with fundus disease and the scarcity of ophthalmologists have brought huge challenges to DR detection.To develop different treatment plans for different stages,and realize the automatic staging of DR is extremely important for the promotion of large-scale DR diagnosis.For the DR automatic staging task,we proposes a deep learning-based diabetic retinopathy staging method(Densely connected Networks based on SE and Modified Inception,SE-MIDNet).Aiming at the problem of inconspicuous lesion features of the data set images,we adjusts the contrast,brightness and color balance of the data set to highlight the lesion features.Aiming at the problem of imbalance in the number of data set categories,we uses an oversampling method to balance the data set.Aiming at the problem that small target features are easily lost when the network is deep due to the different size of the lesion area,we first proposes an improved Inception module,which enables the network to efficiently extract multi-scale features of DR images,thereby enhancing the feature learning ability of the network.The connection method splices the output feature maps of the improved Inception module and sends them to the subsequent layer,so that each improved Inception module can directly obtain the gradient from the loss function and the original input image data,reduce the complexity of the network training process,and realize DR the multi-scale feature reuse of the image enhances the feature representation of small targets.Finally,the channel attention module(Squeeze-and-Excitation,SE)is used to obtain the global information of the feature map on each channel,and rely on this information to perform dynamic non-reflection on each channel.Linear modeling,adaptively distinguish the importance of different features,and weight according to the importance,thereby improving the generalization ability of the network.The experimental results show that for the DR image data set,the improved Inception module can efficiently extract the detailed features of the DR image and improve the accuracy of DR image staging.The accuracy of DR automatic staging reaches 88.24%,the sensitivity reaches 99.43%,and the specificity reaches 97.60%.The designed network structure can realize DR automatic staging and has good generalization ability.